[alfalfa-star]

What

#1599 keeps a baseline of a storage image's guest bytes (images up to 1 MiB) when its layers get their generation, so a write-back can store the GPU results over CPU-stamped blocks whose bytes didn't change. The upload takes it with GuestMemory::ReadCommitted, also on the host-import path, where the upload itself runs on the GPU.

That read goes through the flush hook. When a unit shadow, another image's pending results or a recorded GPU store covers the surface, the hook publishes the shadow and the worker waits for the GPU on every such upload. Stack from PPSA14632, sampled at those waits: Driver::dispatch → ShaderResources::buildComplete / Revalidate → lookupStorageTexture → StorageTexture::Refresh → upload → (generation stamp) → ReadCommitted → FlushForAccess.

The baseline is now taken only when nothing of that kind overlaps the surface (StorageTexture::guestBytesSettled: AnyShadowedOverlaps, AnyPendingOverlaps, Recorder::PendingWriteOverlaps). Otherwise the image has no baseline, and its write-back keeps the CPU's bytes for stamped blocks, the rule before #1599. The TechnicalDebt entry from #1599 now names the condition.

New ctest agc_storage_baseline_settled (agc_driver_recorder_tests --settled-baseline-only): it creates a storage image over a watched host import while a recorded store is pending over part of it, and checks that the upload did not submit the open batch. Without this change it fails ("an upload over a pending recorded store submitted the open batch to read its generation baseline"). agc_unchanged_cpu_stamps (#1599) still passes, so settled surfaces keep the baseline.

Tested

Linux (Arch), RX 9070 XT, RADV, Release:

• Full ctest: 699/699.
• PPSA14632 (LEGO Horizon Adventures), on my test branch (main db343a9 plus perf(agc): sync only the source for DMA_DATA copies into host-only memory #1157, feat(agc): depth planes sampled as 2D arrays, cubes and integer views #425, perf(agc): read a fast-cleared surface's guest bytes only when the import is compared #2699, feat(agc): vertex clip and cull distances #423/feat(agc): render layered draws into 3D color targets #1077 and local stand-ins needed to reach gameplay), steady state per 10 s:

	flush-hook recorded-store syncs	waited	queue 0 direct dispatch, device call
without	~3,400	~4.0 s	356 us
with this PR	~630	~0.3 s	107 us

On its own this barely moves presents (~100 → ~104 per 10 s): the GPU wait moves to DMA_DATA copies, which #2792 addresses.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes

🤖 Generated with Claude Code

[popcoder565]

Native Windows run, since no Windows results were posted. Setup: Windows 10 (19045), RTX 3060 Laptop (NVIDIA 616.92), MinGW-w64 GCC 15.2.0 (WinLibs r7), Release. Head 75b540c2, merged onto main 57f96fd in one build with 4 other new PRs.

• agc_driver_recorder_tests (not in ctest), which holds this PR's test change, passes 10 of 10, built explicitly.
• Full ctest in that build: 711 of 711.

No title was run.

🤖 Addressed by Claude Code

[orbz22]

Checked on Windows with Hollow Knight: Silksong (PPSA12544), the title whose boxed menu letters #1599 fixed.

Setup: Windows 11 Pro (26300), i7-11800H, RTX 3060 Laptop (driver 617.42), WinLibs GCC 15.2.0 r7, Release. main de5e541 + #846 (for the plugins), with and without this PR. Each run used the same scripted input for 250 s (title menu, profile select, New Game, the opening, into Moss Grotto) with APS5_PROFILE_DRAW=1.

• No regression: the menu and profile-select text is intact in frames dumped at 960x540. The opening and Moss Grotto look the same as without the PR.
• Waits: the pending write syncs from FlushForAccess dropped to 0 with this PR. Without it there were 3 to 5 per run, in three runs. They took no measurable time here. The other GPU waits did not change, and most of them come from WaitForKeyWriters.
• agc-driver: write-back drops GPU results of full 64 KiB blocks of 1920x1080 render targets (PPSA12544 backgrounds garbled) #1970: the 8 keeping the CPU's blocks lines appear with and without the PR.

🤖 Generated with Claude Code